Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
NBER WORKING PAPERS SERIES
UNION THREAT EFFECTS AND NONUNIONINDUSTRY WAGE DIFFERENTIALS
David Neumark
Michael L. Wachter
Working Paper No. 4046
NATIONAL BUREAU OF ECONOMIC RESEARCH1050 Massachusetts Avenue
Cambridge, MA 02138April 1992
We thank John Blomquist for research assistance, and William Dickens, Christopher Hanes,David Levine, and Jaewoo Ryoo for helpful comments. Research support was provided by theInstitute for Law and Economics, University of Pennsylvania, This paper is part of NBER'sresearch program in Labor Studies. Any opinions expressed are those of the authors and notthose of the National Bureau of Economic Research.
NBER Working Paper #4046April 1992
UNION THREAT EFFECTS AND NONUNIONINDUSTRY WAGE DIFFERENTIALS
ABSTRACT
We investigate the impact of union strength on changes in nonunion wages and
employment. The prevailing model in this area is the threat model, which predicts that increases
in union strength cause increases in nonunion wages and decreases in nonunion employment.
In testing the threat model, we are also testing two alternatives, the crowding and complements
models. In contrast to the prediction of the threat model, decreases in the percent organized
(reflecting a declining union threat) are associated with increases in the nonunion wage.
Furthermore, increases in union wages appear to decrease, rather than to increase, nonunion
wages. Evidence on the determinants of intra-industry variation in nonunion wage premia is
somewhat more consistent with the crowding model and is strikingly consistent with the
complements model of union and nonunion wage determination. Further evidence on the
determinants of intra-industry variation in nonunion employment is consistent with the
complements model and the threat model; movements in nonunion industry employment are
negatively related to changes in proxies for union strength. Thus, the combined evidence
supports the complements model, but neither the threat model nor the crowding model.
David Neumark Michael L. WatcherAssistant Professor Professor, and Director,Department of Economics Institute for Law and EconomicsUniversity of Pennsylvania Department of EconomicsPhiladelphia, PA 19104 University of Pennsylvaniaand NBER Philadelphia, PA 19104
In the union literature of the past several years, considerable evidence has been
marshalled to support the fact that union employment and overall union strength have
declined dramatically over the 1970s and 1980s, while union wage premiums have
increased just as dramatically (see, e.g., Bell, 1989; Linneman, Wachter, and Carter,
1991; Wachter and Carter, 1990, and Neumark, forthcoming). Little attention has been
paid, however, to the effect of these developments on the nonunion sectors of the
economy. This paper examines these developments, with the goal of providing empirical
tests of alternative models of the effects of union sector developments on nonunion
wages and employment.
The prevailing model of the impact of unionization on the nonunion sector has
been based on the threat model. First tested by Rosen (1969), the threat model predicts
that an increase in union power will cause an increase in nonunion wages. The
nonunion response occurs because nonunion employers calculate that nonunion workers
are less likely to unionize when the nonunion firms are independently responding to
union wage increases. Wage developments in the nonunion sector therefore follow, to
some extent, wage developments in the union sector; for example, an increase in union
wages leads nonunion employers to increase their own wage premium, thereby moving
up their demand curve and laying off workers and reducing output.
In the theoretical and empirical analysis of the union threat model in the existing
literature, the level of analysis is typically the industry rather than the overall economy.
Paralleling this, the union threat model has been offered as an explanation of inter-
industry wage differentials in the nonunion sector (e.g., Dickens, 1986; Dickens and
Katz, 1987a; Krueger and Summers, 1988). Because one goal of our analysis is to test
this explanation of the nonunion wage structure, we also conduct our analysis at the
industry level. Since unions or in Dunlop's terms 'industrial relations systems" (Dunlop,
1944) are not neatly organized along two- or three-digit industry categories, and because
threat effects of any particular union may affect firms in a number of industries ('orbits
of coercive comparison" (Ross, 1948)), we conduct our analysis for both one-digit and
two-digit industry classifications.
There are two principal alternatives to the threat model: the crowding and
complements models. Although our equation specifications are tailored to test the
threat model, we also consider the predictions of the crowding and complements models
for these specifications, to clarify the interpretation of any rejections of the threat model
that may arise.
The "crowding model' is labor supply driven. It is based on the view that
nonunion workers are labor market substitutes for union workers. Higher union wages
and any resulting lower union employment cause an outward labor supply shift in the
nonunion sector. The result is lower wages and higher employment in the nonunion
sector. Although nonunion labor demand may also increase if nonunion output is a
substitute for union output, the prevailing view is that labor supply effects dominate
labor demand effects.
The "complements model," which has received less attention in the literature,
focuses on the labor demand ties between union and nonunion labor within the same
industry. In this model, there are three sectors: a union sector, a "substitute" nonunion
sector that competes with the union sector, and a "complement" nonunion sector that
serves as a production complement to both the union and substitute nonunion sectors.
2
Workers in the complements sector may include nonunion workers employed in the
same establishment as union workers (perhaps in other occupations), or nonunion
workers in nonunion firms that act as suppliers to the union and substitute nonunion
sectors. Lower union sector output would thus reduce the demand for nonunion
complements workers, thereby driving down their wages and employment.
The prevailing empirical literature supports, albeit weakly, the threat model. The
major papers in the area are cross-sectional tests focusing on nonunion wages as the
dependent variable. Little attempt has been made to analyze whether the results
support either the crowding or complements models. Nor have the alternative models
been tested on time-series data, at the industry or the aggregate level.
In this paper we test the threat model in a pooled cross-section time-series data
set, using the period 1973 to 1989, a period during which union wages increased, and
union employment decreased dramatically. We also test the predictions generated by the
other models of the relationships between the union and nonunion sectors.
Finally, we consider an additional dependent variable. The threat model has
typically been analyzed in the context of threat effects on nonunion wages. We extend
the analysis by considering the implications of the threat model, as well as the crowding
and complements models, for nonunion employment as well. We show that by jointly
considering union-nonunion interactions for both wages and employment, we can
provide a sharper differentiation among the competing models.
In general, our results are inconsistent with the union threat model and the
crowding model, and are largely consistent with the complements model. These
3
conclusions come from a number of empirical tests, and hold at both the one-digit and
the two-digit industry level.
Section I discusses the relevant literature and Section II presents the alternative
models. The data are described in Section III and the empirical tests performed in
Section IV. Section V concludes the paper.
I. Literature Review
A formal model of the threat hypothesis, in which nonunion employers pay
above-market-clearing wages to deter unionization, is developed in Dickens (1986). In
this model the determinants of the wage that a nonunion employer must pay to avoid
unionization are the magnitude of supra-competitive rents (which may or may not exist
in the long-run), and the costs of organizing. The large existing literature which ties the
percent organized in an industry to the wages of nonunion workers in cross-sectional
data (e.g., Freeman and Medoff, 1981; Kahn, 1980; Podgursky, 1986), can be interpreted
as testing the validity of this model, on the assumption that the costs of organizing are
negatively related to the percent organized or, equivalently, to the percentage of industry
employment that is unionized.1 Dickens and Katz (1987b) interpret their research tying
nonunion industry wage premia to measures of market power and profitability (in
addition to the percent organized), as conducting indirect tests of the union threat
model, since these variables presumably proxy for industry rents; however, as they
'Rosen (1969) develops this argument in more detail. He suggests that threat effects on nonunionwages will be largest when the demand for union labor is least elastic, since there are then smaller laborflows to the nonunion sector in response to the union wage increase. This elasticity is likely to be smallest(in absolute value) when the industry is relatively more unionized, since there are then fewer substitutionpossibilities, so that the percent organized is positively related to threat effects. But Rosen also argues thatthere may be offsetting factors, perhaps because as the percent organized increases, the remaining nonunionfirms are particularly resistant to unionization.
4
recognize, other theories of non-market-clearing wages would also predict a positive
relationship between rents and wages (e.g., Akerlof, 1982; Lindbeck and Snower, 1988).
Finally, Montgomery (1989) examines the effects of the percent organized and the union
wage premium on employment probabilities of individuals.
The formulation closest to ours is Dickens and Katz (1987a), which conducts an
empirical analysis of the union threat explanation of nonunion industry wage premia.
They use 1983 annual files of the CPS to estimate individual-level wage regressions,
including three-digit Census of Population industry dummy variables.2 In a second-stage
equation they then regress the estimated industry coefficients on a rich set of industry
characteristics, to identify the correlates of industry wage premia.3'4 -
Consistent with the threat hypothesis, Dickens and Katz find that the sign of the
percent unionized variable in the nonunion wage regressions is generally positive. But
they point out that the hypothesis is only weakly supported because the coefficient is not
always positive, and is often statistically insignificant. Some of the other variables do
support the threat model, particularly the positive coefficient on profitability, which they
indicate supports the hypothesis that the higher the potential rents, the higher the wage
(Dickens, 1986). Since industry characteristics are highly correlated, however, Dickens
2The cross-sectional equation includes the measurable human capital of workers, demographiccharacteristics of workers, and controls for state of residence, living in an MSA, and occupation.
3These include: average wages; average income; average demographic characteristics; unemployment andlayoff rates; injury rates; overtime; non-wage compensation; average establishment size; measures of marketpower, concentration, and profitability; other industry characteristics; and, most directly related to the unionthreat explanation, the percent of workers in the industry covered by collective bargaining agreements.
4Dickens and Katz rationalize this 'two-step' approach, arguing that estimating an individual-levelregression with some variables (i.e., industry characteristics) aggregated to the industry level may lead tobiased coefficients because the industry aggregates measure the characteristics of the worker's firm witherror (Dickens and Ross, 1984).
5
and Katz note that it is difficult to sort out the independent influence of the large
number of industry characteristics that they consider. In addition, because the key
percent unionized variable does not provide robust support for the threat model, it is
more difficult to interpret the meaning of the profitability variable.5
As indirect evidence against the threat model, researchers have cited the fact that
the nonunion inter-industry wage structure is quite stable over time and across regions
(e.g., Krueger and Summers 1988; Helwege, 1991). According to these researchers,
given variation in union wage premia across time or regions, a nonunion wage structure
that is highly correlated across time or regions makes it less likely that a strong threat
effect exists. However, the variation in union wage premia that might be expected to
affect the nonunion wage structure is changes in intra-industry union wage premia across
time or regions, not simply changes in overall union wage premia across time or regions.
Consequently, the sharp changes in intra-industry union wage premia in the 1970s and
1980s (Linneman, Wachter and Carter, 1990) provide a natural testing ground for the
union threat explanation of the nonunion wage structure. In particular, we can ask
whether changes have occurred, in nonunion inter-industry wage differentials, that are
related to changes in union threats.
5lndeed, they cite factor analysis results for industry data suggesting that one factor can account formuch of the variation in industry chacacteristics. A related result is suggested in their paper, in which theyshow that two principal components can explain about 50 percent of the covariation in their industry data.
6
II. The Union Threat, Crowding, and Complements Models
Data Set and Dependent Variables
In this paper we construct a panel data set, rather than a cross-sectional data set
as in Dickens and Katz (1987a). The dependent variable--the nonunion wage
differential or premium by industry and year--is estimated from cross-sectional wage
equations using CPS data, including skill and other related control variables. One of the
central independent variables, the union wage premium, is formed in the same fashion.
The method of construction is discussed in Section III.
The nonunion wage premia (denoted w") are then regressed on the proxies for
union strength in a second-stage regression, in which we control for other industry
characteristics with industry-specific dummy variables. The hypothesis is that the greater
the threat effect, the greater the nonunion premium.
In addition, we add a second equation to test the threat model. Besides
estimating an equation where the nonunion wage is the dependent variable, we also
analyze the threat effect in terms of the nonunion industry employment share (denoted
es') as a dependent variable.6 If the union threat increases, it should affect not only
the nonunion wage, but also nonunion employment. We define employment shares as a
percent of total economy employment rather than industry employment. The reason is
twofold. Since the threat model, as well as the crowding and complements models,
suggest spillovers between the union and nonunion sectors in the same industry, a
within-industry variable would be difficult to interpret. In addition, losses in either
6Union and nonunion employment-share variables were estimated directly from the CPS data. Detailsare provided in Section III.
7
union or nonunion employment need not be to the other sector in the same industry, as
shown in Linneman, Wachter, and Carter (1991).
Independent Variables or Measures of Union Threats
The next step is to determine the variables that can serve as measures of the
magnitude of the union threat. The traditional view is to measure union strength by the
percent organized (%org), with a predicted positive coefficient in an equation for
nonunion wage premia (w). This is shown diagrammatically in Figure 1, where a
decrease in %org means a downward shift in the demand for union labor from D° to
and an increase in the demand for nonunion labor from to DnUI.7 The impact of
the decline in union employment is to cause the threat locus, which measures nonunion
firms' perceived threat from unionization, to decline from T to T', allowing w to
decline.8
We expand this framework by introducing the union wage premium as another
measure of the union threat (as in the model in Dickens, 1986), also with a predicted
positive coefficient. In Figure 1, an increase in the union wage (and hence an increase
in the union wage premium) from w to causes a movement up the demand curve
7Since neither sector is a market-clearing sector, and both pay above market wages, employment isdemand constrained. Hence, the supply curves can be omitted.
8The T locus shows the nonunion wage premium that must be paid to deter unionization. It isdownward sloping because as e? is increasing, the threat is decreasing. It shifts when other variables,particularly the union variables described in this section, change.
8
for union labor Di'. This in turn causes the threat locus, T, to shift to 12, causing w"' to
increase and ef' to decrease.9
In addition, we add three additional independent variables that can proxy for the
'costs of organizing. We include two election variables: the percentage of National
Labor Relations Board (NLRB) representation elections won by unions (denoted
%won), and the overall number of representation elections (denoted rep). We also
include the number of unfair labor practice claims against employers (denoted ulp). Our
hypothesis is that the coefficient(s) on the election variables will be positive in the threat
model.1° Specifically, when union certification elections are low and when unions are
losing many of the elections, the relative cost of remaining nonunion decreases, and T
shifts down to T3.
On the other hand, when the number of unfair labor practice charges against
management is high, unions themselves are facing increased costs of organizing workers
or maintaining their union status. U/p charges are high when management's opposition
to unions is high and unions contest the legality of management's activities.
Management opposition in this form allows nonunion wages to be lower; an increase in
u/p shifts the threat locus, T, downwards. Hence, in the threat model the hypothesis is
that the coefficient of ulp will be negative.
Our data formulation has the advantage that the statistical experiment is
conducted on a data set in which there are sharp changes in variables that are plausibly
9Some difficulties posed by using this variable in an equation where the nonunion wage w°°1' is thedependent variable are discussed below. The union wage variable, however, is pivotal in the equationwhere the nonunion employment share is the dependent variable.
'°Notationally, we refer to the two election variables %won and rep, when discussed jointly, as elect.
9
related to 'the strength of the union threat effect. The specifications also include
industry dummy variables. The intent is to have the dummy variables capture the other
industry characteristics considered by Dickens and Katz (1987a), so that we can identify
union threat effects from our estimated coefficients. Using industry fixed effects avoids
the problem found in cross-sectional studies of multicollinearity among industry
characteristics. However, to identify union threat effects, the omitted characteristics that
are correlated with the threat variables must be fixed over time. This is the maintained
assumption in this paper; given the sharp changes in union strength over the sample
period, it may not be far from the truth.
The union strength proxies are treated as exogenous, because we do not believe
that there are identifying restrictions available for explicit treatment of the endogeneity
question. Denoting the industry fixed effect I, and subscripting the variables by industry
(i) and year (t), we estimate equations of the form
(1) w" = a + %orgB1 + wuPj32 + elect1B3 + u1p134 + ly +
Our specification of threat effects on the nonunion employment share parallels
the specification of equation (1). All of the variables from (1) are included in (2) with
the exception of %org. Since the dependent variable is the nonunion employment share,
using the percent organized as an independent variable could result in a spurious
negative correlation. This is the case even though our dependent variable has total
economy employment as the divisor while the independent variable would have industry
employment as the divisor.
10
Denoting the nonunion employment share as es", the equation for the nonunion
employment share is
(2) esv =a' + w.B'2 + elect113'3 + ulp1B'4 + ly' +
Since the nonunion sector is presumed to be acting like the union sector, in
response to increases in union strength or the union wage it moves up its labor demand
curve, thereby reducing employment. Hence, the threat model implies that the signs of
the union threat proxies are reversed in equation (2); the nonunion employment share
moves inversely with the nonunion wage premium. The effects on nonunion wages and
employment that are implied by the threat model are summarized in the first panel of
Table 1.
Th Crowding Model
The crowding model focuses on the implications of the labor supply spillover
from the union sector, where employment is restricted by a wage premium, to a market-
clearing nonunion sector. Whereas the nonunion sector "acts like the union sector" in
the threat model, the nonunion sector reacts competitively in the crowding model. Thus,
the responses of nonunion wages and employment to an increase in the union wage
premium, w, are the reverse of those in the threat model. This, as well as the other
differences and similarities between the two theories can be seen by comparing Figures
1 and 2.
Turning to %org, a decrease in %org causes D to shift to DU!, thus causing S to
shift out to S. The result is a decline in w9. Note that in the crowding model, the
11
traditional assumption is that labor supply spillovers dominate labor demand spillovers.
Hence, although the nonunion demand curve shifts out, the maintained assumption in
the model is that the nonunion supply curve shifts out by more. Thus, we show only the
latter shift. The resulting decline in w"° is the same as in the threat model, but for a
different reason; this is the only variable for which the two theories generate the same
prediction in equation (1).
The predicted effects of the NLRB variables are also different from those of the
threat model. One has to be careful, however, as to the interpretation given these
variables in the crowding model. The NLRB variables are introduced to serve the
primary mission of the paper, testing the threat model. With the exception of %won,
they might not be introduced as likely variables in a crowding model, were we testing
that model independently of the threat model. But we want to consider the implications
for the coefficients of these variables in equations (1) and (2), under the crowding
model (and the complements model), in order to interpret results that may be
inconsistent with the threat model.
%won is a useful variable because it measures actual or potential shifts between
the union and nonunion sector. For example, a decline in %won means that unions are
losing elections and thus D is shifting inward to D', and and S° are shifting out to
D' and as shown in Figure 2. For w, the effect is a wash, but es clearly
increases. In general %won operates in a similar fashion to %org, but has a greater
demand effect.
We interpret the NLRB variables rep and ulp as measuring costs of union labor
not reflected in w, and hence operating in a similar fashion to w'', although changes in
12
these variables result in shifts of Du, rather than shifts along it.1' The lower the
number of union elections and the greater the allegations of unfair labor practices by
management, the weaker is the union, and hence the lower are labor costs in the union
sector. Hence a decrease in rep or an increase in ulp causes an outward shift in the
demand for union labor, from DU to DU4, leading nonunion labor supply to shift in to
S'4. These shifts cause es' to decrease, and thus w to increase.
The results for the crowding model are summarized in the second panel of Table
1. In closing, we note that the crowding model is sometimes specified to include a
possible shift in labor demand in the nonunion sector in the opposite direction of the
employment change in the union sector. Including both the spillover of demand and
supply from the union sector, the final effects on w and esare ambiguous, depending
on the magnitudes of the shifts in labor supply and demand. In the existing literature, it
is usually assumed that the supply shifts dominate. For example, Kahn (1978) contrasts
the threat and crowding models by the different impacts that union wage increases have
on nonunion wages (positive for the threat model, and negative for the crowding model).
We make the same assumption here.
The Complements Model
Finally, we consider a model in which union and nonunion labor are complements
in production. This is a more complex model, since it requires the depiction of two
nonunion sectors. It does, however, capture a well recognized multiplier effect, whereby
"For example, consider two industries that have identical union wage premia, but the first industry hasmore representation elections and weaker management opposition to unions. We might expect unions in thisindustry to extract higher non-wage concessions from management, resulting in a higher cost of union labor.
13
a decline in union employment will spill over into a decline in nonunion employment
that works along with, or as a complement to, union labor.
The model is depicted in Figure 3. The nonunion substitute sector in this figure
works in the same manner as did the nonunion sector in Figure 2. The difference is that
we postulate the existence of a nonunion sector which follows the level of activity in the
union sector.
In this model, a decline in %org causes the usual decrease in union employment
and increase in demand for nonunion labor that acts as a substitute. In this case, the
outward shift of S° to S'1 causes an increase in es"' and a decline in Although not
drawn, DS may also shift outward, which would moderate the wage decline. But as in
the crowding model, it is assumed that the dominant impact of the union sector on the
nonunion substitute sector is supply shifts. More importantly, however, the overall
composite wage across the union and nonunion substitute sectors declines. The result is
an increase in demand in the complements sector. Hence D' shifts to D1. We assume
that developments in the nonunion complement sector dominate the overall nonunion
sector. Thus, the prediction is that both ef and w'°" will increase.
Moving on to the Wt variable, an increase in w' in the complements model acts
as a depressant on activity in the union sector without the same corresponding uptick in
nonunion substitute sector employment as was the case with %org variable. The result
is that D' declines to D2, causing both es and w to decline.
As noted above %won operates in a similar fashion to %org. Hence a decrease
in %won is like a decline in %org, causing both es and w to increase. Similarly rep
14
and ulp act like w"'. Hence a decrease in rep and an increase in ulp causes both es'" and
w""" to increase.
The predicted coefficient signs of the complements model are presented in the
third panel of Table 1. Scrutinizing Table 1 reveals the advantage of considering both
equation (2) (for employment) as well as equation (1) (for wages) if we are to
differentiate among the three models. Equation (1) provides sharply contrasting
predictions between the threat model, on the one hand, and the crowding and
complements models, on the other; but only the coefficient of %org, B, differentiates
between the crowding and complements models. Equation (2), in contrast, provides
sharply contrasting predictions between the crowding model, on the one hand, and the
threat and complements models, on the other.
Test of the Models Based on Union Sector Adjustments
We also carry out an additional test, which is a more "direct test of the extent to
which nonunion industry wage premia estimated from cross-section wage regressions
measure threat effects within the industry. This test considers the feedback of the size
of the threat effect on the responsiveness of union employment to changes in the union
wage. Specifically, where threat effects are strong, the decline in union employment in
response to an increase in the union wage should be relatively weak. That is, union
employment is better protected against increases in union wage premia when the
nonunion sector validates the union premium by matching it with an increase in its own
premium.
15
We use the nonunion industry wage premia w to index threat effects, and test
whether the data are consistent with these premia reflecting threat effects. Denoting the
union share of total employment esu and the union wage premium w't', we estimate
equations of the form12
(3) es1 = ô + w°'1X + + (WUP5 x w"'"11)O + +
In general, union employment should respond negatively to an increase in the
union wage, so we should find X < 0. But threat effects should temper this response.
Consequently, if the nonunion industry wage premia sv"" index threat effects, then we
should find 0 > 0; that is, the decline in union employment in response to an increase in
the union wage should be less negative in industries in which there are strong threat
effects. The sign of , the coefficient of W"', is ambiguous, since w can be viewed as
the price of another input along with union labor. If nonunion and union labor are
complements, then i should be negative; can also be negative if they are substitutes,
but the scale effect of an increase in w dominates.
As before, we also consider the predictions of the crowding and complements
models for the sign of 0 in equation (3). The crowding model yields the same prediction
as the threat model for this regression. If union and nonunion labor are substitutes, then
the union labor demand response to an increase in the cost of union labor should be
smaller if accompanied by an increase in the cost of nonunion labor (i.e., a higher value
of wut), and vice versa. Thus, the crowding model also predicts that 0 > 0.
°We use the union share of total employment, rather than the union share of industry employment, tominimize the endogeneity problem. The union share of industry employment is significantly influenced bynonunion industry employment, which may in turn influence the right-hand-side variable w.
16
The complements model, however, yields opposite predictions from the threat
and crowding models. A high value of the nonunion wage premium corresponds to a
high price of a complementary input. Thus, a hike in union labor costs coupled with a
high nonunion wage premium results in a larger drop in demand for union labor. That
is, the complements model predicts that we should find 0 < 0, in contrast to the threat
model. These results are summarized in the bottom portion of Table 1.
III. Data on Nonunion Wage Premia, Union Wage Premia,and Union Strength
Our estimates of union and nonunion wage premia come from cross-sectional
wage equations estimated for individual years from CPS data. Specifically, we estimated
cross-sectional log wage equations for each year 1973-1989, with separate equations for
white men, black men, and women.13 These equations include a union status
(membership) dummy variable, and a full set of interactions of all variables, including
the industry dummy variables, with the union status dummy variable.'4'5 To focus on
competitive market effects of union wages, we exclude government workers from our
sample. In addition, we exclude managers, professionals, and technical workers. In part
because many of these workers are not covered by the National Labor Relations Act,
'3i982 is omitted because no data were collected on union membership.
'4Details of the concordance between 1970 and 1980 SIC codes are available from the authors.
°ln addition to these variables, we included: nine one-digit occupation dummy variables (with a bridgebetween the 1970 and 1980 SOC codes); linear and quadratic schooling; linear and quadratic potentialexperience; dummy variables for four regions; the MSA unemployment rate; dummy variables for threeMSA sizes; and dummy variables for married, spouse present, and overtime (based on usual hours worked).
17
these occupations have low rates of unionization, and changes in their wages and
employment are likely to be least affected by developments in the union sector.
The coefficients of the non-interacted industry dummy variables estimate the
nonunion industry wage premia.16 Our estimate of the within-industry union premium
is the sum of the coefficients of the industry-union interaction and the union status
dummy variable. In all cases we estimated the wage premia for the entire sample by
weighting (by industry and union employment) the coefficients estimated from the
separate regressions by race and sex.
Our measure of percent organized is constructed from the same CPS data,
described immediately above, used in estimating the wage premium variables. The
percentage of representation elections won by unions, the number of representation
elections held, and the number of unfair labor practice charges against employers are
from NLRB Annual Reports.'7
We constructed a data Set at both the one- and two-digit industry levels. Because
the NLRB threat proxies are available only at the one-digit level, the two-digit industry
data are used only to explore the more limited set of specifications using the union and
nonunion industry wage premia and employment shares, and the percent organized.
In Table 2 we provide summary statistics for the one-digit industry data set,
reporting mean levels and yearly changes for each of these variables, by industry. The
ft"Industry wage premia are estimated relative to services for the one-digit analysis, and relative to
business services for the two-digit analysis. The regression results are of course insensitive to whichindustry is omitted. We replicated many of our results defining industry wage premia instead relative to theaverage nonunion worker (weighting by the representation of workers across industries). The conclusionswere unchanged.
'7Details are provided in the footnotes to Table 2.
18
yearly changes are related to the data used to estimate the regressions, although fixed-
effects are used rather than first differences. The data confirm the now well-known
increase in the within-industry union premium (w') in many industries, coupled with the
decline in percent organized (%org). For the NLRB variables, the percentage of
elections won by unions (%won) and the number of representation elections (rep)
decline for nearly all industries, while the number of unfair labor practice charges (ulp)
increases for two-thirds of the industries.
In focusing on this time-series cross-sectional data set an interesting question is
raised: was the threat effect posed by unions increasing or decreasing during the 1973-
1989 period? On the one hand, the threat effect as measured by the percent organized
and the other NLRB measures was declining. As the most well-known development in
the union sector, the prevailing view is likely to be that the threat effect was declining.
On the other hand, the wage differential between union and nonunion workers within
industries, which is related to the potential gains from unionization, tended to increase
over this period, in manufacturing as well as other industries (Linneman, Wachter, and
Carter, 1990, and our Table 2). Since the weight one should give to the opposing factors
is unclear, the overall direction of the threat effect over the 1973 to 1989 sample period
is indeterminate. But, as we shall show, the question is moot. Although the nonunion
sector did respond systematically to developments in the union sector over the sample
period, the changes are not those predicted by the threat model.
19
IV. Empirical Results
Potential Biases
We have identified four sources of bias in the estimates of our equations that
may affect the conclusions to which the estimates lead. First, in equation (1) we use
as a measure of the potential wage gain from unionization, while wis the dependent
variable. The v"' variable is obtained from the wage regression as the sum of the
coefficients on the union variable and the union-industry interaction variables. The w
variable, on the other hand, is the coefficient on the industry dummy variable. To the
extent that the covariation between w and w' comes from changes in nonunion
industry wage premia that do not similarly affect union workers in the industry (perhaps
because they are locked into contracts), a negative correlation arises between the
dependent variable and the within-industry union premium that does not reflect causal
influences running from the latter variable to the former. On the other hand, variation
in w generated by union-specific forces will impact the union coefficient or the union-
industry coefficient without creating this spurious negative correlation. Because of this
potential problem, when we discuss the empirical results for equation (1), the nonunion
wage premium equation, we focus most on the specifications excluding the within-
industry union premium, w''.
A second potential problem with the within-industry union premium is that it may
reflect the influence of underlying (possibly unobservable) union threat effects, because
this measure may reflect the equilibrium after the effects of union threats are
assimilated into nonunion wages. For example, the within-industry premium may be
large in an industry precisely because unobserved union threat effects are weak, leading
20
to negative bias. Including industry dummy variables should help, however, by
controlling for unobserved industry characteristics that are fixed over time. This also
leads us to focus on specifications excluding the within-industry union premium.
Third, in specifications excluding w", the percent organized (%org) is a key
variable of interest. The estimated coefficient of this variable may, however, be upward
biased by endogeneity. Omitted factors that shift up the nonunion premium, w, may
lead to declines in nonunion employment, and hence increases in %org; this generates a
positive association between the residual in equation (1) and %org. The implication of
this is that the results are biased against finding a negative coefficient on 91oorg, which,
as Table 1 shows, would be evidence in favor of the complements model. Since much of
our evidence ultimately points toward the complements model, this potential source of
bias only strengthens our results.18
Fourth, we have so far ignored the potential effects of industry-specific demand
shocks (stemming, perhaps from import competition) that affect both union and
nonunion labor. These could generate potentially important biases if, as seems
plausible, union wages are less flexible than nonunion wages, with union employment
consequently more flexible. In this case a downward industry demand shock, for
example, will result in a decrease in the nonunion premium ("),a decrease in the
nonunion employment share (esou), an increase in the within-industry union premium
(wuP), and a decrease in the percent organized (%org). Looking at Table 1, we can see
that in equation (1), the equation for the nonunion wage premia, this biases the
We do not believe that we have identifying information on the basis of which to correct for theendogeneity bias.
21
coefficient of w downward, and the coefficient of the percent organized upward.
Together, these biases will tend to favor the crowding model. In equation (2), the
equation for the nonunion employment share, this biases the coefficient of w"
downwards, creating a bias against the crowding model. To control for industry-specific
demand shocks, we introduce the percentage change in the GNP share accounted for by
each industry to attempt to control for shifts in industry-specific demand for labor,
although we recognize that this variable is not unambiguously exogenous. To account
for differences across industries in cyclical fluctuations in this variable, we estimated
separate regressions, for each industry, of the GNP share on the aggregate
unemployment rate and a post-1976 dummy variable for the change in accounting
methods, and used the residual in the regression estimates of equations (1) and (2).1920
Results
In Table 3 we report results from regression estimates of equation (1), in which
the nonunion industry wage differential is the dependent variable. In columns (1)-(7)
we use the one-digit industry data.2' In column (1) we present the standard model of
the union threat effect. The negative coefficient on the percent organized variable is
the opposite of that predicted by the threat model, since the estimated coefficient says
The GNP-share data are from CITIBASE for 1973-1976, and from the Survey of Current Business forthe later years.
20To explore the role of aggregate demand shocks, we experimented with the incJusion of year dummyvariables in our equations, but their coefficients were never statistically significant. Here, in contrast, we arereferring to demand shocks that affect specific industries.
21UnJess otherwise specified, statements regarding statistical significance refer to live-percent significancelevels. For regression coefficients, the statements refer to two-sided tests.
22
that the greater the union strength in organizing the industry, the lower the wage in the
remaining nonunion sector.
In column (2), the NLRB variables are added to the equation. The predictions of
the threat model are that the two election variables are positively related to the
nonunion wage premium, while the number of unfair labor practice claims is negatively
related. The percent organized variable remains largely unchanged, while the NLRB
variables present a mixed picture. The coefficients of the elections variables are
positive, consistent with the threat model, although neither is statistically significant.n
The unfair labor practice variable, on the other hand, has the opposite sign from that
predicted bythe threat model, although it, too, is statistically insignificant. In columns
(3) and (4) we replace percent organized with the within-industry union premium, first
alone and then along with the NLRB variables. In column (3) the coefficient of the
within-industry union premium is negative and strongly significant.
This result rejects the threat model, but the conclusion must be tempered by the
potential for a spurious negative correlation for reasons noted above.
When the NLRB variables are added, in column (4), the evidence against the
threat model is more one-sided. The coefficients of both election variables are negative,
We recognize, though, that the relationship between the percent organized and the union threat may beambiguous, as in Rosen (1969). However, for two reasons we have a strong expectation of a positiverelationship. First of all, while the difficulty of unionizing may become quite strong once very high levels ofunionization are reached, this hardly seems to characterize unionization rates in the U.S. in the sampleperiod. Second, much of the individual-level evidence on the union threat hypothesis looks for and findspositive relationships between individuals' wages and the percent of their industry unionized, although thisrelationship may be present only for certain demographic groups (Kahn, 1980), or in large firms (Podgursky,1986), or it may be rather weak statistically (Freeman and Medoff, 1981).
nThere was no evidence of severe multicollinearity between the two elections variables; in general, thecoefficient estimates and standard errors of the election variables were little changed when one or the otherwas omitted from the equation.
23
and the number of elections variable is statistically significant. The coefficient of the
unfair labor practices variable remains positive but not statistically significant;
furthermore, the three NLRB variables are jointly significant. The within-industry union
premium coefficient remains negative and significant.
In column (5) we reintroduce the percent organized variable. The results again
are almost entirely the opposite of those predicted by the threat model. Both the
percent organized and within-industry premium variables have statistically significant
negative coefficients. The only NLRB variable that is significant is the unfair labor
practice variable, and this also has the positive sign noted above.
In column (6) we add a variable capturing changes in the share of GNP
contributed by each industry, to the specification from column (2). As discussed at the
beginning of this section, this is intended to control for biases induced by industry-
specific demand shocks. The estimated coefficients are little changed. The same is true
when we add back in the within-industry union premium, in column (7).
In columns (8)-(1O), we report results for the subset of the specifications that can
be estimated using data at the two-digit industry level of disaggregation. The greater
disaggregation implies that our estimates of industry wage premia and employment
shares are estimated from smaller cells, leading to greater measurement error bias and
less precise coefficient estimates. The results in columns (8)-(1O) nonetheless parallel
closely the estimates of the corresponding one-digit specifications. As in the estimates at
the one-digit level, the estimates of the %org and W' coefficients are negative and
statistically significant.
24
*Besides providing strong results against the union threat model, the pattern of
results in Table 3 provides mixed support for the crowding model. Remember that in
the crowding model, the nonunion sector reacts competitively and, assuming that union
and nonunion labor are substitutes, generally in the opposite direction of the union
threat model. The crowding model successfully predicts the negative coefficient on w"'
and the positive coefficient (in some of the specifications) on ulp, at the one-digit level,
and the corresponding negative coefficient on w°, at the two-digit level. These
coefficients were incorrectly predicted by the threat model. In terms of w', the
crowding model predicts that the higher the union wage premium, the more crowding
will occur in the nonunion sector, hence the lower is w. In terms of ulp, the crowding
model predicts that the higher the number of unfair labor practice allegations, the
weaker is union power, reflecting more active management opposition, and hence, the
lower is the cost of union labor. This, in turn, means a higher w reflecting the weaker
union sector.
The percent organized variable is the one exception to the rule that the crowding
model has the opposite prediction from the threat model. For this variable, the direct
supply effects are important and probably outweigh the cost of labor interpretation
assigned to the other threat variables. Viewed in this fashion, the crowding model, like
the threat model, does not predict the negative coefficient on %org. That is, the
crowding model predicts that the lower the percent organized, the more crowding there
is in the nonunion sector, and hence the lower should be w. In fact, however, the
estimates in Table 3 indicate that lower %org leads to higher w'.
25
The complements model, however, is more successful in predicting the pattern of
results found in Table 3. Remember that in the complements model, nonunion labor
that is complementary to union labor follows the level of economic activity in the union
sector, and this complementary nonunion labor dominates the nonunion sector overall.
The implication is that, with respect to regression equation (1), the complements model
has the reverse predictions from the threat model. In the threat model, increases in
union power cause the nonunion sector to mimic the union sector, moving up its demand
curve and increasing w. In the complements model, however, nonunion labor that is a
production complement to union labor responds to higher labor costs by decreasing its
wage (and reducing its employment share).
Viewed in this fashion, the complements model successfully predicts the negative
coefficient on w; the higher w', the higher the cost of union labor and thus the lower
the level of economic activity in the union sector. The less activity in the nonunion
sector that is a complement to the union sector, the lower w'. Similarly, the
complements model correctly predicts the positive sign on the unfair labor practice
variable. In terms of u!p, the higher the level of management opposition, the lower the
cost of union labor and thus the higher WUP. Finally, the negative sign on %org is
consistent with complementarity between union labor and nonunion labor, overall.
In summary, the result of estimating equation (1) is a set of coefficients that are
almost entirely contrary to the predictions of the threat model, are mixed with respect to
the predictions of the crowding model, but confirm the predictions of the complements
model. This conclusion holds at both the one- and two-digit industry levels.
26
To gain further evidence on the threat model, and to further distinguish between
the alternative models, we next turn to estimates of equation (2), in which the nonunion
employment share of total economy employment, es, is the dependent variable.
Results for these regressions at the one- and two-digit levels are reported in Table 4.
This equation has the same specification as equation (1) with the exception that the
%org variable is excluded. Although the dependent variable, es, is a share of total
economy employment, it is likely to be spuriously negatively correlated with %org (the
union share of industry employment).
In each of the specifications estimated in Table 4, all of the coefficients are
statistically significant, except for %won (and ulp in columns (2) and (3)). Focusing on
the statistically significant coefficients, with respect to the threat model, the results of
Table 4 are largely supportive. Increases in w cause decreases in es, and increases in
ulp or decreases in rep cause increases in es. Given the results in Table 3, however,
the result with respect to '' is suspect. The reason that es should fall in the threat
model is that W" has increased in response to increases in w. However, we know
from Table 3 that this is not supported by the data. Hence, the decrease in es as a
consequence of increases in w cannot support the threat modeL
For the corresponding analysis at the two-digit level, only one specification can be
estimated, the regression of es on wu (and the industry dummy variables). The
regression is reported in column (5) of Table 4. The coefficient of w is negative, but is
not statistically significant. This is consistent with the one-digit results in the table,
although it is not, by itself, informative.
27
The crowding model is rejected by the results in Table 4. Increases in w should
cause additional crowding in the nonunion sector, hence increasing es. But the
coefficiept on w'' is negative. Similarly the crowding model incorrectly signs the ulp and
representation elections variables in Table 4.
The complements model, however, is supported by the results of Table 4. In the
complements model the negative coefficient on the w variable is predicted. Since
increases in w'' cause a decline in economic activity in the union sector, they also cause
a decrease in economic activity in the nonunion complements sector, hence the decline
in es"'. Similarly the positive sign on ulp, which is an indirect measure of the cost of
union labor, is predicted by the complements model, as is the negative sign on the
representation elections variable.
Overall, the results for the employment share regressions suggest that increases in
union strength are associated with declines in nonunion industry employment. By itself,
this finding is consistent with either the threat model or the complements model, but not
the crowding model. However, only the complements model is consistent with both sets
of findings from the employment and wage regressions. Furthermore, it is reasonable to
expect that production complementarities between union and nonunion labor are weaker
at a more disaggregated level of analysis, since the disaggregation is more likely to sort
the complementary types of labor into different industries. Because of this, we view the
two-digit industry results for the wage and employment regressions as particularly
compelling evidence in favor of the complements model.
Finally, we turn to estimates of equation (3), intended to test directly whether
nonunion industry wage premia are positively related to union threat effects. Table 5
28
reports results at the one- and two-digit level. To summarize the test, the coefficient of
the within-industry union premium should be negative, reflecting movements along the
labor demand curve for union labor. If nonunion industry wage premia are positively
related to threat effects, then the interaction coefficient should be positive, as strong
threat effects diminish the substitutability between union and nonunion workers. A
positive interaction coefficient would also be consistent with the crowding model. On
the other hand, a negative coefficient on the interaction variable would be consistent
with the complements model, indicating that increases in union wages costs coupled with
increases in nonunion wage costs lead to relatively larger declines in union employment.
The estimates in column (1) are consistent with the complements model. The coefficient
of the interaction variable (-.276) is negative and statistically significant. Furthermore,
the coefficient of the within-industry union premium is negative, as expected, and the
negative coefficient of the nonunion industry wage premium is consistent with
complementarity. In column (2) we explore whether the coefficient of the interaction
variable is being identified from cross-industry variation in the responsiveness of union
employment to union wages (i.e., union labor demand elasticities). To do this we add a
set of variables interacting the within-industry union wage premium with a set of
industry dummy variables.25 The signs of the coefficients are unchanged, but the
coefficient on the interaction variable is no longer statistically significant. Columns (3)
Wedefine the nonunion industry wage premium used to construct the interaction variable as deviationsabout its mean. Consequently, the partial derivative of union employment with respect to the within-industry union wage premium, evaluated at the means, is given by the coefficient of the within-industryunion wage premium.
25We define the industry dummy variables relative to their sample means so that the within-industryunion wage premium coefficient still measures the partial derivative, evaluated at the sample means.
29
and (4) repeat this analysis at the two-digit level, with the same results; the signs of the
point estimates are consistent with the complements model, although the evidence is
statistically significant only in column (3).
Overall, the evidence in Table 5 provides additional support for the complements
model; as we suggested earlier, we find the evidence in favor of the complements model
at the two-digit level particularly compelling. Based on these findings, we can more
decisively conclude that the results provide no evidence that nonunion industry wage
premia are correlated with the strength of the union threat within industries.
V. Conclusions
In this paper we investigated the impact of union strength on changes in nonunion
wages and employment. The prevailing model in this area is the threat model, which
predicts that increases in union strength cause increases in nonunion wages and
decreases in nonunion employment. In testing the threat model, we are also testing two
alternatives, the crowding and complements models.
With respect to nonunion industry wage differentials, the statistical experiment
does not ask whether one can explain the 'entire' nonunion industry wage structure with
the union threat effect model, but only asks whether the threat effect model explains
some of the variation in this wage structure over time. The answer appears to be no. In
contrast to the prediction of the threat model, decreases in the percent organized
(reflecting a declining union threat) are associated with increases in the nonunion wage.
Furthermore, increases in union wages appear to decrease, rather than to increase,
nonunion wages. Evidence on the determinants of intra-industry variation in nonunion
30
wage premia is somewhat more consistent with the crowding model and is strikingly
consistent with the complements model of union and nonunion wage determination.
Given that the union threat model does not explain intra-industry variation in nonunion
industry wage premia over time, it seems an unlikely candidate for an explanation of the
cross-sectional variation in nonunion industry wage premia.
Further evidence on the determinants of intra-industry variation in nonunion
employment is consistent with the complements and the threat model; movements in
nonunion industry employment are negatively related to changes in proxies for union
strength.
Thus, the combined evidence supports the complements model, but neither the
threat model nor the crowding model. Evidence from a third test, based on the
responsiveness of union employment to union wage changes (again, within industry),
strengthens this finding, as it is consistent with the complements model, but neither the
threat model nor the crowding model.
In testing the model on a cross-section, time-series data base from 1973-1989, we
are also able to explore how the extraordinary changes in the union sector during this
period have affected the nonunion sector. In the context of the threat model, increases
in the union wage premium in many industries, coupled with sharp declines in union
employment shares, raise the possibility that the strength of the union threat in different
industries varied little over this period. Thus, the threat model is not necessarily
inconsistent with the observed stability of the nonunion industry wage structure over the
31
same period. While our evidence rejects the threat model in favor of the
complements model, it still suggests that union employment shares and union wage
premia have offsetting influences on nonunion wages; but the effects have the opposite
signs from those predicted by the threat model. Thus, it is possible that union sector
developments do influence the nonunion industry wage structure, but these influences
are masked in the 1970s and 1980s because of offsetting trends in the union sector.
"Krueger and Summers (1988) reports correlations around .91 between industry wage differentialsestimated using CPS data in 1974, 1979, and 1984.
32
References
Akerlof, G.A. 1982. "Labor Contracts as Partial Gift Exchange.' Quarterly Journal ofEconomics 91(4): 543-69.
Bell, L.A. 1989. 'Union Concessions in the 1980s." Federal Reserve Bank of New YorkQuarterly Review Summer: 44-58.
Dickens, W.T. 1986. "Wages, Employment and the Threat of Collective Action byWorkers.' Mimeograph, University of California, Berkeley.
Dickens, W.T. and L.F. Katz. 1987a. "Inter-Industry Wage Differences and IndustryCharacteristics." In K. Lang and J.S. Leonard, Eds., Unemployment and the Structure ofLabor Markets (New York, NY: Basil Blackwell), 48-89.
1987b. "Inter-Industry Wage Differences and Theories of Wage Determination."NBER Working Paper No. 2271.
Dickens, W.T. and B.A. Ross. 1984. "Consistent Estimation Using Data from More thanOne Sample." NBER Technical Working Paper No. 33.
Dunlop, John T. 1944. Wage Determination Under Trade Unions (New York, NY:Macmillan).
Freeman, R.B. 1986. 'The Effect of the Union Wage Differential on ManagementOpposition and Union Organizing Success." American Economic Review 76(2): 92-6.
Freeman, R.B. and J.L. Medoff. 1981. 'The Impact of the Percentage Organized onUnion and Nonunion Wages." The Review of Economics and Statistics LXIII(4): 561-72.
Heiwege, J. 1991. "Sectoral Shifts and Interindustry Wage Differentials." Mimeograph,Federal Reserve Board.
Kahn, L.M. 1978. 'The Effect of Unions on the Earnings of Nonunion Workers."Industrial and Labor Relations Review 31(2): 205-16.
1980. "Union Spillover Effects on Organized [sic] Labor Markets." Journal ofHuman Resources 15(1): 87-98.
Krueger, A.B. and L.H. Summers. 1988. "Efficiency Wages and the Inter-Industry WageStructure." Econometrica 56(2): 259-94.
Lindbeck, A. and D.J. Snower. 1988. The Insider-Outsider Theory of Employment andUnemployment (Cambridge, MA: The MIT Press).
Linneman, P.D., M.L. Wachter, and W.H. Carter. 1990. 'Evaluating the Evidence onUnion Employment and Wages." Industrial and Labor Relations Review 44(1): 34-53.
Montgomery, E. 1989. "Employment and Unemployment Effects of Unions." Journal ofLabor Economics 7(2): 170-90.
Neumark, D. "Declining Union Strength and Labor Cost Inflation in the 1980s."Forthcoming in Industrial Relations.
Podgursky, M. 1986. 'Unions, Establishment Size, and Intra-Industry Threat Effects."Industrial and Labor Relations Review 39(2): 277-84.
Rosen, S. 1969. 'Trade Union Power, Threat Effects and the Extent of Organization."Review of Economic Studies 36(2), No. 106: 185-96.
Ross, Arthur M. 1948. Trade Union Wage Policy (Berkeley, CA: University of CaliforniaPress).
Wachter, M.L. and W.H. Carter. 1989. "Norm Shifts in Union Wages: Will 1989 Be aReplay of 1969?" Brookings Papers on Economic Activity 2: 233-64.
up
Figure 1
Union Threat Hypothesis
1. Decrease in Dora (D — V, D — D')
2. Increase in it to e°
3. Decrease In Doss, rep,
increase Is alp
esDecreases Increases (by assiaiptlsn)
Increases Decreases
— 1' Decreases Increases
5sp2 - -
Union Sector Nonunion Sector
Effect
T —.
"UPI
"up
1. Decrease in %org
2. Increase in a' to a'
3. Decrease in %aon
Figure 2
Crowding (Supply) Alternative
—D", C —
C • SM
D' — D', C —
Note: We asstana shifts in C dominate shifts in C, end therefore generalLy do not show the Latter.
"up wnup
'pup
a a
Union Sector
as
Nonunion Sector
CDecreases
Decreases
No change
on
Increases (by asaueptim)
Increases
Increases
4. Decrease in rep, increase in oLp D' —. D', C C' Increases Decreases
Figure 3
Complements (Demand) Alternative
Union Sector Nonunion sector (Comp.) Nonunion Sector (Subet.)
eC1. Decrease in Zere -. V, s — C', increases Increases (by assteçtien)
-. D'
2. Increase In a' to a° D' — D Decreases Decreases
3. Decrease is Deeti V -. D,D D, Increases IncreasesC -. r,D -.4. Decrease In rep, increase In sip D' — D', D -. D Increases Increases
este: We assune that mevee,ents In eases end esçlsytnent of nsrsoion cmp(ee,ent Labor dominate the nsnunisn setter. We assuseshuts In D dominate shifts in C, and therefore do net show the Latter.
coups 'suDs
e ea eJ' es eJ°
Table 1
Summary of Coefficient Predictions from Alternative Models
gions 1 and 2: Nonunion wage (1') Nonunion emolovinent (2')
Threat Model
%org + not included
+
%won +
rep +
ulp +
Crowding Model
%org + not included
+
%won no change
rep +
ulp +
Complements Model
%org not included
—
%won — -rep - -
ulp + +
Eanation 3: Union employment (3')
Threat ModelA -
0 +
Crowding ModelA -
w' x w 0 +
Complements ModelA -
x W" 0
Table 2
Descriptive Statistics by One—Digit Industry
Nonunion Within-industry Nonnioo Union Percent Unfair lar
wage union wage eaploylent eaployzent Percent elentiors Reprenentation practice chargespresiua prealea shares shares organized won by noOns elactiors' against eeployers'
(1) (2) (3) (4) (5) (6) (7) (8)
Construction .00 .36 .05 .03 .35 .49 .05 .09
(.01) (.01) (.001) (.002) (.02) (.01) 1,01) (.01)
Wining .34 .16 .01 .01 .40 .43 .01 .04(.02) (.02) (.0004) (.0005) (.03) (.01) (.001) (.004)
Nanufacturing— .17 .17 .12 .08 .40 .43 .20 .63durable (.01) (.01) (.001) (.01) (.02) (.01) (.02) (.33)
Manufacturing— .12 .15 .09 .04 .33 .42 .13 .37mn—durable (.01) (.02) (.001) (.003) (.01) (.01) (.02) (.02)
t'raosgartation .22 .19 .04 .05 .52 .48 .11 .36cowaunications, and (.02) (.02) (.001) (.002) (.02) (.01) (.01) (.01)piblic utilities
8bolesale trade .15 .15 .05 .01 .13 .42 .06 .13(.01) (.02) (.002) (.0004) (.00) (.01) (.01) (.01)
Retail trade —.06 .29 .13 .02 .12 .41 .10 .25
(.01) (.01) (.01) (.001) (.01) (.01) (.01) (01)
Finance, insurance, .12 .05 .07 .003 .05 .53 .02 .03and real estate (.02) (.03) (.003) (.0001) (.002) (.02) (.001) (.003)
Services ... .14 .17 .03 .16 .54 .17 .39(.01) (.004) (.001) (.004) (.01) (.01) (.02)
Table 2 (continued)
thanoen
10000ion itbin-indostry Ronunion Union Percent Unfair 1ar
sage union sage eaployaent eaploysent Percent electiose Representation practice charges
preai presini share shar? organizeS son by unions elections' against eiployero'
(1) (2) (3) (4) (5) (6) (7) (8)
I:Cnnstruction .005 —.001 .001 —.001 —.012 .002 .007 .005
(.006) (.010) (.011) (.001) (.009) (.014) (.005) (.007)
1ining .012 —.004 .0002 —.0002 —.021 —.012 —.0002 .004
(.031) (.037) (.0003) (.0003) (.011) (.020) (.001) (.003)
Xanufactoriog— .003 .003 —.0000] —.005 —.014 —.005 —.016 —.002
durable (.012) (.022) (.002) (.003) (.005) (.005) (.009) (.020)
Oannfacturing— .00] .001 —.00005 —.002 —.011 —.003 —.010 —.0001
non—durable (.013) (.025) (.001) (.001) (.005) (.009) (.006) (.014)
Transgartatioo, .001 .003 .001 —.001 —.012 —.007 —.00] .005
onuoicatiorw, and (.019) (.027) (.001) (.001) (.009) (.009) (.005) (.009)
piblic utilities
holesa1e trade .002 .003 .001 —.0002 —.006 —.009 —.004 .00]
(.018) (.031) (.001) (.0001) (.005) (.013) (.004) (.007)
Retail trade .0003 .004 .003 —.001 —.006 —.006 —.007 —.002
(.007) (.014) (.002) (.0004) (.003) (.009) (.003) (.007)
Finance, inscrance, .005 —.002 .001 —.00003 —.001 .000 —.0007 .001
and real estate (.021) (.058) (.004) (.002) (.002) (.019) (.001) .002)
Services .,. .003 .003 .000] -.001 -.002 -.0003 .014
(.017) (.003) (.001) (.004) (.020) (.009) (.011)
1. 1leaos are reportad, with standard errors is parentheses.
2. Share of total eaplnysent.3. The representation elections variable is the nuiber of cases receiveS in the fiscal year stealing fros petitions fileS by a laher organizatiow or
esployee seeking an election for detaraioatinn of a collective bargaining representative. The ousber of elections is divideS by 10000, The unfairlaher practice variable is the nuaber of charges fileS by labar organizations or eaployee5 against esployers. the nuaber of charges is divideS by
10,000.
Table 3
Nonunion Wage Regressions, 1973—1989, Weighted Least SquaresDependent Variable: Nonunion Industry Wage Premium1
One-Dinit Results Two-Digit Results(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Percent organized —.273 —.328 —.248 —.327 —.236 -.107 ... —.106
(.040) (.046) (.038) (.048) (.039) (.023) (.019)
Percentage of —.034 .011 .019 .014
elections eon (.039) (.035) (.031) (.040) (.032)by unions
Number of -.126 .004 .104 -.024
representation (.063) (.055) (.052) (.069) (.057)
elections/10,000
Unfair labor 045 .079 .046 .077
practice cbarges (.044) (.040) (.035) (.044) (.035)
against employers
/10,000
Within—industry —.452 -.476 —.404 ... —.411 ... —.377 -.376union premium (.052) (.051) (.046) (.046) (.029) (.028)
SN? sbare2 -.044 —.558
(.569) (.454)
.964 .965 .969 .971 .978 .965 .978 .938 .953 .956
1. Standard errors are reported in parentheses. There are 144 observations for tbe one-digit aoalysin, and480 observations for tbe two—digit analysis. Industry dummy variables are included. Observations areeeigbted by number of sbservations for year from ebicb micro—level coefficients were estimated. The within-industry union and nonunion pre.iums are estimated from log wage regressions using the outgoing rototiongroup annual files of the CPS for 1983—1989, and tbe Nay files for 1973-1981. Regressiuns care estimatedseparately by race esd ses, omitting government workers, managers esd prufessinoals, and tochsical workers.The premiums were then calculated es weigbted averages of the premiums estimated from the separate wageregressions. Other variables included in tbe individual—level wage regressions were: one—digit industrydummy variables (construction, mining, durable manufacturing, nondurable manufacturing, transportation,finance, insurance and real estate, retail trade, wbolesale trade, and services); nine one—digit occupationdummy variables (with a bridge between the 1970 and 1980 SOC codes); linear and guadratic schooling; linearand guadratic potential esperience; dummy variables for four regions; the NSA unemployment rate; and dummyvariables for three NSA sizes; and dummy variables for married, spouse present, overtime (based on anualbourn worked).2. Thin is the residual from a regression estimated for eacb industry of the SN? share of output produced bythe industry on an intercept, the aggregate civilian unemployment rate, and a post—1976 dummy variable tocapture the cbange in accounting methods used is the numbers reported in the Surveo of Current Resiness.
Table 4
Nonunion Employment Share Regressions, 1973—1989,Weighted Least Squares
Dependent Variable: Nonunion IndustryEmployment/Total Employment'
00e-Dioit Results No-Digit Results(1) (2) (3) (4) (5)
Within—industry —.033 ... —.046 —.040 — .000
union premium (.017) (.016) (.015) (.006)
Percentage of 010 .019 .011
elections woo (.011) (.011) (.010)
by unions
Nurberof ... —.082 —.090 —.057
representation (.017) (.017) (.016)elections/l0,000
Onfuir labor ... .017 .022 .022
practice charges (.013) (.013) (.011)against employers/10,000
asP share2 823
(.143)
.979 .982 .983 .986 .953
1. Standard errors are reported in parentheses. There are 144 observations f or the one—digit analysis, and 480 observations for the two-digit analysis. Industry dummy vuriableaare included. Observations are weighted by number of observations for year from whichmicro—level coefficients were estimated. See footnotes to Table 3 f or rare details.
Table 5
Employment Share Response Regressions, 1973—1989,Weighted Least Squares
Dependent Variable: Union Share of Total E]nployment
Within—industryunion wage premium (wa')
Nonunion industrywage premium (w)
Within—industryunion wage premium (w°)x nonunionindustry wagepremium (w)'
Includes industry dummyx within—industryunion wage premiuminteractions
R'
One—Digit Results
(1)
—.078(.025)
—.084(.036)
—.276(.135)
(2)
— .022(.071)-.095(.052)
—.186(.185)
Two—Digit Results
(3) (4)
—.007 .002(.003) (.017)
—.011 —.010(.004) (.005)—.039 —.030(.015) (.024)
No Yes
.805 .817
No Yes
.872 .874
1. Standard errors are reported in parentheses. industry danny variables are included. There are 144 observations for theone—digit analysis, and 400 observations for the two—digit analysis. Observations are weighted by saber of observations foryear fro. abich iicrs—level coefficients were estijated. See footnotes to Table 3 for lore details.2. This is included as deviations about its lean, no that the coefficient of the within-industry union wage preniun neasuresthe partial derivative of the eiploynent share variable with respect to the within-industry union preniuw, evaluated at theleans.